Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
tags:
|
| 4 |
+
- security-research
|
| 5 |
+
- picklescan-bypass
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
# Picklescan Scanner Bypass PoC — codecs.open LFI
|
| 9 |
+
|
| 10 |
+
⚠️ **This is a security research PoC. Do NOT load untrusted model files.**
|
| 11 |
+
|
| 12 |
+
## Summary
|
| 13 |
+
|
| 14 |
+
This repository demonstrates a bypass in [picklescan](https://github.com/protectai/picklescan) (v0.0.22), the scanner used by HuggingFace Hub to detect malicious pickle payloads in model files.
|
| 15 |
+
|
| 16 |
+
The file `picklescan-bypass-poc.joblib` contains a payload using `codecs.open("/etc/passwd", "r")` which:
|
| 17 |
+
- **Passes picklescan with 0 issues detected** (classified as "suspicious" not "dangerous")
|
| 18 |
+
- **Successfully reads /etc/passwd** when loaded via `joblib.load()`
|
| 19 |
+
|
| 20 |
+
## Scan Result
|
| 21 |
+
|
| 22 |
+
```
|
| 23 |
+
Issues: 0
|
| 24 |
+
Infected: 0
|
| 25 |
+
Globals: [Global(module='codecs', name='open', safety=<SafetyLevel.Suspicious>)]
|
| 26 |
+
```
|
| 27 |
+
|
| 28 |
+
## Affected Versions
|
| 29 |
+
|
| 30 |
+
- picklescan ≤ 0.0.22
|
| 31 |
+
|
| 32 |
+
## Root Cause
|
| 33 |
+
|
| 34 |
+
`codecs.open` is not in picklescan's `_unsafe_globals` blocklist. It is only classified as `Suspicious`, which does NOT increment `issues_count` or `infected_files`, and does NOT block model loading.
|
| 35 |
+
|
| 36 |
+
## Responsible Disclosure
|
| 37 |
+
|
| 38 |
+
This vulnerability has been reported via Huntr.com's Model File Vulnerability Program.
|